基于数据挖掘技术的高光谱土壤质地分类研究

【目的】 寻找红壤地区不同土壤质地类型的Vis-NIR光谱反射规律,通过光谱对土壤质地类别进行快速、准确的预测。【方法】 以江西省奉新县北部为研究区,245个土壤样本为研究对象,在国际制土壤质地4组和12级两种分类标准下,首先分析不同土壤质地类型的光谱反射率,然后采用9种数学变换方法和5种机器学习算法相互组合的数据挖掘模型,进行土壤质地的分类研究,最后对建模准确度最高的混淆矩阵和预测结果三角坐标分布图进行分析。【结果】 （1）不同土壤质地之间的光谱反射率存在较多的交叉重叠现象,土壤质地与光谱反射率之间的规律较为复杂;（2）分数阶导数变换是整数阶导数的扩展,有助于土壤质地的分类,但原始光谱数据具有更加丰富的特征信息,更适合进行土壤质地分类建模;（3）在对非均衡数据集建模时,集成学习方法和神经网络方法都是不错的选择;（4）较难通过模型去区分土壤质地分界线附近的类别,其中在4组分类标准下最容易被预测错误成黏壤土组,在12级分类标准下最容易被预测错误成黏壤土和壤质黏土这两种土壤质地类型;（5）在4组分类标准中,进行归一化处理和MLP模型组合取得了0.68的最高预测准确度,其中黏壤土组的预测准确度能达到0.84;再细分到12级分类后,分类效果最佳的组合来自于原始数据和MLP模型,其中壤质黏土分类准确度达到了0.89。【结论】 本研究结果可为南方红壤地区通过高光谱数据进行土壤质地分类提供参考依据。

Soil Texture Classification of Hyperspectral Based on Data Mining Technology

【Objective】 The aim of this study was to find the reflection law of Vis-NIR spectra of different soil texture types in red soil region, and to quickly and accurately predict the soil texture type by the spectrum. 【Method】 Taking the north of Fengxin County in Jiangxi Province as the research area, 245 soil samples were taken as the research objects. Under the 4 groups and 12 levels of international soil texture classification standards, the spectral reflectance of different soil texture types was analyzed first, then the data mining models combining 9 mathematical transformation methods and 5 machine learning algorithms were used to classify the soil texture, and finally analysis of the confusion matrix with the highest modeling accuracy and the triangular coordinate distribution map of prediction results. 【Result】 (1) There were many overlaps and overlaps in the spectral reflectance between different soil textures, and the law between the soil texture and the spectral reflectance was more complicated. (2) Fractional derivative transformation was an extension of the integer derivative, which was helpful for the classification of soil texture, but the original spectral data had more abundant feature information and was more suitable for the classification of soil texture. (3) Both ensemble learning methods and neural network methods were good choices when modeling unbalanced data sets. (4) It was difficult to distinguish the categories near the boundary of soil texture by using the model. Among them, clay loam group was the most likely to be predicted wrongly under the four classification standards, and clay loam and loamy clay were the two most likely to be predicted wrongly under the 12 classification standards. (5) Among the four groups of classification standards, the highest prediction accuracy (at 0.68) was obtained by the combination of normalization treatment and MLP model, and the prediction accuracy of clay loam group could reach 0.84. After subdivision to 12 levels classification, the best classification result came from combination of original data and MLP model, and the classification accuracy of loamy clay was 0.89. 【Conclusion】 The results of this study could provide a reference for soil texture classification by using hyperspectral data.